A region-specific multi-task data-driven framework for deciphering and mitigating indirect carbon emissions in wastewater treatment plants.

Indirect carbon (C) emissions from wastewater treatment plants (WWTPs), primarily driven by energy consumption (EC) and chemical consumption (CC), are challenging to quantify. While data-driven models offer a promising solution, its application is often limited by data constraints or incomplete coverage of C accounting. Thus, this study developed a regional data-driven framework and selected Hubei Province as a case study using the 2024 monthly data for 177 WWTPs. Spatially, population-dense and economically developed areas such as the biggest Wuhan city showed higher EC and CC. The Gradient boosted decision trees (GBDT) model excelled in EC prediction (R2test = 0.86, RMSEtest = 0.101 kWh/m3) better than random forest (RF), XGBoost, and CatBoost. MultiCCNet performed best for CC (R2test = 0.72 ± 0.04, RMSEtest = 0.007 ± 0.005 kg/m3) compared with MultiGBDT and MMoE. Post-hoc interpretability analysis identified optimal operational ranges [such as load rate (75-85%), volume (above 2.0 million m3), influent suspended solids (0-100 mg/L), influent ammonia nitrogen (8-12 mg/L), influent chemical oxygen demand (above 125 mg/L)] for low EC and key levers including chemical types, load rate, and wastewater quality characteristics for mitigating CC. The models demonstrated robustness with validated data in 2025, and scenario analysis showed potential emission reductions of 59.0% for EC and 97.4% for CC. This study provided a scientific framework for precise C accounting and mitigation in WWTPs, which could be extended to support and assist other regions as well.